Benchmarks? about powerboxes HOT 2 CLOSED

Hi !
Thanks for taking the time to look this :) and for the numpy nms function
The discrepancy between the benchmarks in the readme indeed came from the number of trials which were too low in your case (number of cpu cores doesn't matter since neither numpy nor opencv seem to use multithreading here)

I updated your code (thanks again) to plot error bars, and to increase number of trials. I will integrate this in the colab notebook which holds the benchmarks visualizations. Here are the results I obtain on my laptop with a 12th Gen Intel(R) Core(TM) i7-1250U cpu

One thing to note is that powerboxes nms scales particularly badly with the number of boxes compared to opencv, which in turn performs badly on a small number of boxes. The logic in opencv is quite different than what I implemented (actually it's adaptive nms). I will look into it and see if there's a way to take best of both approaches.

from powerboxes.

You are right about the threading. Numpy version is basically using vectorization to speedup the calculations. In fact, I tried Cython version of the same thing and yet numpy's native python implementation is twice as faster. Using jit (e.g. Jax or Numba) doesn't help either as the code cannot be parallelized efficiently.

As for OpenCV, it does uses adaptive NMS, however, implementation is unclear as I haven't seen the C++ code yet. Another thing is that OpenCV version of the code be further optimized using OPENCL or CUDA acceleration which is such a pain to setup in python.

What you have here (powerboxes) is a very well implemented and very efficient. Especially without threading, the performance is quite amazing.

from powerboxes.